feat: 新增 Agent CLI — Web 版 AI 运维 Agent（参考 opsxcli 移植）
后端引擎 backend/pkg/agentcli/（纯引擎库，零业务依赖，剥离 CLI/TUI 外壳）： - llm: OpenAI 兼容客户端，复用 model.LLMConfig - tools: ssh_execute（复用 pkg/remoteexec）+ analyze_output + 风险评估黑名单 - safety: 三级安全模式 + 五级风险 + WebSocket 异步确认 - prompt: 五层 System Prompt - core: ReAct 16 轮循环 + 自适应检查点 + 循环检测 - evolver: 四层记忆 + 进化循环，data/agentcli/ 文件存储 集成层： - controller/agent_cli.go: WebSocket 控制器（事件流推流 + 高危确认 + 会话持久化） - model/agent_cli.go: agent_cli_sessions/messages 表 + AutoMigrate - 路由 /ws/agent-cli + /api/v1/aiops/agent-cli/sessions - seed 权限 aiops:agent-cli / aiops:llm-config 前端： - AgentCLI.tsx: 对话流 + 工具块 + 高危确认 + 四模块 tips 面板 - LLMConfig.tsx: 大模型设置独立页 - api/agentCli.ts + Sidebar/App 菜单路由 文档： - docs/agent/ 11 篇（opsxcli 同名文档适配改写 + 00-README 映射总览） 其他： - CLAUDE.md 补充 agent-cli 模块说明 - 修正 .gitignore 根级二进制规则（前导 /，避免误伤 frontend/src/components/aiops/） Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
attackerandClaude committed
1 parent
f2069fe3f0
commit
92a2247390
37 files changed
+6449
-6
No files matched your search
@@ -0,0 +1,436 @@
|
||||
// Package core 实现 Agent CLI 的 ReAct 核心引擎。 // Logan.Li
|
||||
// 设计参考 opsxcli internal/agent/core，剥离 TUI（Bubble Tea），改用 EventCallback 推送事件流
|
||||
// （由 controller 序列化为 WebSocket 消息推给前端 + 持久化到 DB）。
|
||||
// 核心算法保留：16 轮 ReAct 主循环、自适应检查点、循环检测（同参>3 次拒绝）、连续失败追踪、Evolver 集成。
|
||||
package core
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"fmt"
|
||||
"strings"
|
||||
"sync"
|
||||
"time"
|
||||
|
||||
"ezops/pkg/agentcli/evolver"
|
||||
"ezops/pkg/agentcli/llm"
|
||||
"ezops/pkg/agentcli/prompt"
|
||||
"ezops/pkg/agentcli/safety"
|
||||
"ezops/pkg/agentcli/tools"
|
||||
)
|
||||
|
||||
// EventType 事件类型（驱动前端渲染与 DB 持久化） // Logan.Li
|
||||
type EventType string
|
||||
|
||||
const (
|
||||
EventThought EventType = "thought" // AI 思考/推理文本
|
||||
EventToolCall EventType = "tool_call" // 即将执行的工具（名/参数/风险）
|
||||
EventToolResult EventType = "tool_result" // 工具执行结果（输出/成败/耗时）
|
||||
EventProgress EventType = "progress" // 轮次进度 + token
|
||||
EventDone EventType = "done" // 任务完成（最终答案）
|
||||
EventError EventType = "error" // 错误（含达到最大轮次）
|
||||
)
|
||||
|
||||
// Event 引擎产生的事件（controller 序列化后通过 WebSocket 推送） // Logan.Li
|
||||
type Event struct {
|
||||
Type EventType `json:"type"`
|
||||
Round int `json:"round,omitempty"`
|
||||
Thought string `json:"thought,omitempty"`
|
||||
ToolName string `json:"tool_name,omitempty"`
|
||||
ToolArgs map[string]interface{} `json:"tool_args,omitempty"`
|
||||
Risk string `json:"risk,omitempty"`
|
||||
Output string `json:"output,omitempty"`
|
||||
Success bool `json:"success,omitempty"`
|
||||
Duration string `json:"duration,omitempty"`
|
||||
Tokens int `json:"tokens,omitempty"`
|
||||
Answer string `json:"answer,omitempty"`
|
||||
Message string `json:"message,omitempty"`
|
||||
}
|
||||
|
||||
// EventCallback 事件回调（controller 注入） // Logan.Li
|
||||
type EventCallback func(Event)
|
||||
|
||||
// Config Agent 运行配置 // Logan.Li
|
||||
type Config struct {
|
||||
MaxIterations int
|
||||
Temperature float64
|
||||
ToolTimeout time.Duration
|
||||
MaxTokens int
|
||||
SafetyMode safety.Mode
|
||||
OutputMaxLength int
|
||||
MaxContextTokens int
|
||||
MemoryDir string // Evolver 记忆目录
|
||||
}
|
||||
|
||||
// DefaultConfig 默认配置 // Logan.Li
|
||||
func DefaultConfig() *Config {
|
||||
return &Config{
|
||||
MaxIterations: 16,
|
||||
Temperature: 0.3,
|
||||
ToolTimeout: 120 * time.Second,
|
||||
MaxTokens: 4096,
|
||||
SafetyMode: safety.ModeBalanced,
|
||||
OutputMaxLength: 10000,
|
||||
MaxContextTokens: 8000,
|
||||
MemoryDir: "data/agentcli",
|
||||
}
|
||||
}
|
||||
|
||||
// Agent ReAct 核心引擎 // Logan.Li
|
||||
type Agent struct {
|
||||
llm llm.Client
|
||||
registry *tools.Registry
|
||||
config *Config
|
||||
safetyCtl *safety.Controller
|
||||
evolver *evolver.EvolverEngine
|
||||
callback EventCallback
|
||||
totalTokens int
|
||||
mu sync.Mutex
|
||||
}
|
||||
|
||||
// NewAgent 创建引擎。evolver 为 nil 时降级（无自学习）。 // Logan.Li
|
||||
func NewAgent(client llm.Client, registry *tools.Registry, cfg *Config, sc *safety.Controller, ev *evolver.EvolverEngine) *Agent {
|
||||
if cfg == nil {
|
||||
cfg = DefaultConfig()
|
||||
}
|
||||
if sc == nil {
|
||||
sc = safety.NewController(cfg.SafetyMode)
|
||||
}
|
||||
if cfg.MaxContextTokens <= 0 {
|
||||
cfg.MaxContextTokens = 8000
|
||||
}
|
||||
return &Agent{llm: client, registry: registry, config: cfg, safetyCtl: sc, evolver: ev}
|
||||
}
|
||||
|
||||
// SetCallback 注入事件回调 // Logan.Li
|
||||
func (a *Agent) SetCallback(cb EventCallback) { a.callback = cb }
|
||||
|
||||
// emit 安全触发回调 // Logan.Li
|
||||
func (a *Agent) emit(e Event) {
|
||||
if a.callback != nil {
|
||||
a.callback(e)
|
||||
}
|
||||
}
|
||||
|
||||
// Run 执行单次任务（ReAct 循环 + Evolver 集成） // Logan.Li
|
||||
func (a *Agent) Run(ctx context.Context, query, serverInfo string) (*tools.Result, error) {
|
||||
a.totalTokens = 0
|
||||
start := time.Now()
|
||||
|
||||
// Evolver PREDICT：获取历史经验注入 system prompt
|
||||
evolveCtx := ""
|
||||
if a.evolver != nil {
|
||||
evolveCtx = a.evolver.GetContextForPrompt(query)
|
||||
}
|
||||
|
||||
// 工具描述
|
||||
toolDescs := a.buildToolDescriptions()
|
||||
systemPrompt := prompt.BuildSystemPrompt(serverInfo, toolDescs, evolveCtx)
|
||||
|
||||
messages := []llm.Message{
|
||||
{Role: "system", Content: systemPrompt},
|
||||
{Role: "user", Content: query},
|
||||
}
|
||||
llmTools := a.registry.ToLLMTools()
|
||||
|
||||
toolCallHistory := make(map[string]int) // 循环检测
|
||||
var toolRecords []evolver.ToolCallRecord
|
||||
var consecutiveFailures int
|
||||
var totalToolCalls int
|
||||
|
||||
for iter := 0; iter < a.config.MaxIterations; iter++ {
|
||||
// 检查中断
|
||||
if err := ctx.Err(); err != nil {
|
||||
return nil, err
|
||||
}
|
||||
|
||||
a.emit(Event{Type: EventProgress, Round: iter + 1})
|
||||
|
||||
// 自适应检查点
|
||||
if hint := a.checkAdaptiveCheckpoint(iter, toolCallHistory, consecutiveFailures, totalToolCalls); hint != "" {
|
||||
messages = append(messages, llm.Message{Role: "user", Content: hint})
|
||||
}
|
||||
|
||||
// 裁剪历史
|
||||
trimmed := a.trimMessages(messages)
|
||||
|
||||
// 调用 LLM
|
||||
resp, err := a.llm.Complete(ctx, &llm.CompletionRequest{
|
||||
Messages: trimmed,
|
||||
Tools: llmTools,
|
||||
Temperature: a.config.Temperature,
|
||||
MaxTokens: a.config.MaxTokens,
|
||||
})
|
||||
if err != nil {
|
||||
a.emit(Event{Type: EventError, Message: "LLM 调用失败: " + err.Error()})
|
||||
return nil, fmt.Errorf("LLM调用失败: %w", err)
|
||||
}
|
||||
|
||||
a.totalTokens += resp.Usage.TotalTokens
|
||||
a.emit(Event{Type: EventProgress, Tokens: a.totalTokens})
|
||||
|
||||
// 推送思考文本
|
||||
if strings.TrimSpace(resp.Message.Content) != "" {
|
||||
a.emit(Event{Type: EventThought, Round: iter + 1, Thought: resp.Message.Content})
|
||||
}
|
||||
messages = append(messages, resp.Message)
|
||||
|
||||
// 无工具调用 → 最终答案
|
||||
if len(resp.Message.ToolCalls) == 0 {
|
||||
answer := resp.Message.Content
|
||||
a.emit(Event{Type: EventDone, Answer: answer, Tokens: a.totalTokens})
|
||||
a.triggerEvolve(ctx, query, serverInfo, answer, toolRecords, true, start)
|
||||
return &tools.Result{Success: true, Output: answer}, nil
|
||||
}
|
||||
|
||||
// 处理工具调用
|
||||
observations, records, failInc := a.processToolCalls(ctx, resp.Message.ToolCalls, toolCallHistory, iter+1)
|
||||
toolRecords = append(toolRecords, records...)
|
||||
messages = append(messages, observations...)
|
||||
totalToolCalls += len(resp.Message.ToolCalls)
|
||||
consecutiveFailures = a.updateConsecutiveFailures(observations, consecutiveFailures) + failInc
|
||||
}
|
||||
|
||||
// 达到最大轮次
|
||||
msg := fmt.Sprintf("达到最大迭代次数(%d)，任务未能完成。可能任务过于复杂或需要更多步骤。", a.config.MaxIterations)
|
||||
a.emit(Event{Type: EventError, Message: msg, Tokens: a.totalTokens})
|
||||
a.triggerEvolve(ctx, query, serverInfo, "", toolRecords, false, start)
|
||||
return nil, fmt.Errorf("%s", msg)
|
||||
}
|
||||
|
||||
// processToolCalls 处理工具调用，返回 observation 消息 + 记录 // Logan.Li
|
||||
func (a *Agent) processToolCalls(ctx context.Context, calls []llm.ToolCall, history map[string]int, round int) ([]llm.Message, []evolver.ToolCallRecord, int) {
|
||||
obs := make([]llm.Message, 0, len(calls))
|
||||
records := make([]evolver.ToolCallRecord, 0, len(calls))
|
||||
var failInc int
|
||||
|
||||
for _, tc := range calls {
|
||||
key := tc.Function.Name + ":" + tc.Function.Arguments
|
||||
history[key]++
|
||||
|
||||
// 循环检测
|
||||
if history[key] > 3 {
|
||||
obs = append(obs, llm.Message{Role: "tool", ToolCallID: tc.ID, Name: tc.Function.Name,
|
||||
Content: "错误: 检测到循环调用，相同参数已调用超过3次，请换一种方式"})
|
||||
continue
|
||||
}
|
||||
|
||||
tool, err := a.registry.Get(tc.Function.Name)
|
||||
if err != nil {
|
||||
obs = append(obs, llm.Message{Role: "tool", ToolCallID: tc.ID, Name: tc.Function.Name,
|
||||
Content: fmt.Sprintf("错误: 工具 '%s' 未找到", tc.Function.Name)})
|
||||
continue
|
||||
}
|
||||
|
||||
var args map[string]interface{}
|
||||
if err := json.Unmarshal([]byte(tc.Function.Arguments), &args); err != nil {
|
||||
obs = append(obs, llm.Message{Role: "tool", ToolCallID: tc.ID, Name: tc.Function.Name,
|
||||
Content: "错误: 参数解析失败: " + err.Error()})
|
||||
continue
|
||||
}
|
||||
|
||||
risk := a.toolRisk(tool, args)
|
||||
a.emit(Event{Type: EventToolCall, Round: round, ToolName: tc.Function.Name, ToolArgs: args, Risk: risk.Display()})
|
||||
|
||||
// 安全检查（高危触发 confirmFn，由 controller 异步确认）
|
||||
approved, err := a.safetyCtl.Check(tool, args)
|
||||
if err != nil {
|
||||
obs = append(obs, llm.Message{Role: "tool", ToolCallID: tc.ID, Name: tc.Function.Name,
|
||||
Content: "安全检查失败: " + err.Error()})
|
||||
continue
|
||||
}
|
||||
if !approved {
|
||||
obs = append(obs, llm.Message{Role: "tool", ToolCallID: tc.ID, Name: tc.Function.Name,
|
||||
Content: "操作已被用户拒绝执行"})
|
||||
a.emit(Event{Type: EventToolResult, Round: round, ToolName: tc.Function.Name,
|
||||
Output: "（用户拒绝执行）", Success: false})
|
||||
continue
|
||||
}
|
||||
|
||||
// 执行
|
||||
toolStart := time.Now()
|
||||
toolCtx, cancel := context.WithTimeout(ctx, a.config.ToolTimeout)
|
||||
result, execErr := tool.Execute(toolCtx, args)
|
||||
cancel()
|
||||
dur := time.Since(toolStart)
|
||||
|
||||
var output, errMsg string
|
||||
success := false
|
||||
if result != nil {
|
||||
output, errMsg, success = result.Output, result.Error, result.Success
|
||||
}
|
||||
if execErr != nil {
|
||||
errMsg = execErr.Error()
|
||||
success = false
|
||||
}
|
||||
|
||||
// 输出截断
|
||||
if len(output) > a.config.OutputMaxLength {
|
||||
output = output[:a.config.OutputMaxLength] +
|
||||
fmt.Sprintf("\n[输出已截断，原始 %d 字符]", len(output))
|
||||
}
|
||||
|
||||
a.emit(Event{Type: EventToolResult, Round: round, ToolName: tc.Function.Name,
|
||||
Output: output, Success: success, Duration: dur.Truncate(time.Millisecond).String()})
|
||||
|
||||
records = append(records, evolver.ToolCallRecord{
|
||||
ToolName: tc.Function.Name, Args: args, Output: output,
|
||||
Duration: dur, Success: success, RiskLevel: int(risk),
|
||||
})
|
||||
|
||||
// 构建 observation
|
||||
var content string
|
||||
switch {
|
||||
case execErr != nil:
|
||||
content = "执行错误: " + execErr.Error()
|
||||
failInc++
|
||||
case !success:
|
||||
content = fmt.Sprintf("执行失败: %s\n输出: %s", errMsg, output)
|
||||
failInc++
|
||||
default:
|
||||
content = output
|
||||
}
|
||||
obs = append(obs, llm.Message{Role: "tool", ToolCallID: tc.ID, Name: tc.Function.Name, Content: content})
|
||||
}
|
||||
return obs, records, failInc
|
||||
}
|
||||
|
||||
// triggerEvolve 异步触发进化（带超时 + panic 恢复，绝不阻塞主循环） // Logan.Li
|
||||
func (a *Agent) triggerEvolve(ctx context.Context, query, serverHost, answer string, records []evolver.ToolCallRecord, success bool, start time.Time) {
|
||||
if a.evolver == nil {
|
||||
return
|
||||
}
|
||||
exec := &evolver.TaskExecution{
|
||||
Query: query, ToolCalls: records, TotalSteps: len(records),
|
||||
TotalTokens: a.totalTokens, Duration: time.Since(start),
|
||||
Success: success, FinalAnswer: answer, Timestamp: time.Now(), ServerHost: serverHost,
|
||||
}
|
||||
go func() {
|
||||
defer func() { _ = recover() }()
|
||||
eCtx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
|
||||
defer cancel()
|
||||
_ = a.evolver.Evolve(eCtx, exec)
|
||||
}()
|
||||
}
|
||||
|
||||
// buildToolDescriptions 工具描述文本 // Logan.Li
|
||||
func (a *Agent) buildToolDescriptions() string {
|
||||
var sb strings.Builder
|
||||
for _, t := range a.registry.List() {
|
||||
sb.WriteString(fmt.Sprintf("- %s: %s\n", t.Name(), t.Description()))
|
||||
}
|
||||
return sb.String()
|
||||
}
|
||||
|
||||
// toolRisk 取工具风险（动态优先） // Logan.Li
|
||||
func (a *Agent) toolRisk(tool tools.Tool, args map[string]interface{}) tools.RiskLevel {
|
||||
if dt, ok := tool.(tools.DynamicRiskTool); ok {
|
||||
return dt.RiskLevelForArgs(args)
|
||||
}
|
||||
return tool.RiskLevel()
|
||||
}
|
||||
|
||||
// checkAdaptiveCheckpoint 自适应检查点（iter 3/7/11 注入引导） // Logan.Li
|
||||
func (a *Agent) checkAdaptiveCheckpoint(iter int, history map[string]int, consecFail, totalCalls int) string {
|
||||
switch iter {
|
||||
case 3:
|
||||
repeat := 0
|
||||
for _, c := range history {
|
||||
if c > 1 {
|
||||
repeat += c - 1
|
||||
}
|
||||
}
|
||||
if totalCalls > 0 && (float64(repeat)/float64(totalCalls) > 0.5 || consecFail >= 3) {
|
||||
return "【系统提示】检测到重复使用相同工具或连续失败，请尝试换一种思路或不同命令。"
|
||||
}
|
||||
case 7:
|
||||
return "【系统提示】已执行到中段，请总结进展。信息足够就直接给出最终答案，否则说明剩余步骤。"
|
||||
case 11:
|
||||
return fmt.Sprintf("【系统提示】仅剩 %d 步，请立即整理已有信息给出最终答案，不要再调用新工具。", a.config.MaxIterations-iter-1)
|
||||
}
|
||||
return ""
|
||||
}
|
||||
|
||||
// updateConsecutiveFailures 连续失败计数 // Logan.Li
|
||||
func (a *Agent) updateConsecutiveFailures(obs []llm.Message, current int) int {
|
||||
errs := 0
|
||||
for _, o := range obs {
|
||||
c := o.Content
|
||||
if strings.HasPrefix(c, "错误") || strings.HasPrefix(c, "失败") || strings.HasPrefix(c, "执行错误") || strings.HasPrefix(c, "执行失败") {
|
||||
errs++
|
||||
}
|
||||
}
|
||||
if errs == 0 {
|
||||
return 0
|
||||
}
|
||||
return current + errs
|
||||
}
|
||||
|
||||
// trimMessages 裁剪历史（保留 system + 最近消息，token 不超 90%） // Logan.Li
|
||||
func (a *Agent) trimMessages(msgs []llm.Message) []llm.Message {
|
||||
if len(msgs) <= 1 {
|
||||
return msgs
|
||||
}
|
||||
maxT := a.config.MaxContextTokens * 9 / 10
|
||||
sysT := estimateTokens(msgs[0].Content)
|
||||
kept := make([]llm.Message, 0, len(msgs))
|
||||
keptT := 0
|
||||
for i := len(msgs) - 1; i >= 1; i-- {
|
||||
t := estimateTokens(msgs[i].Content)
|
||||
if sysT+keptT+t > maxT {
|
||||
break
|
||||
}
|
||||
keptT += t
|
||||
kept = append([]llm.Message{msgs[i]}, kept...)
|
||||
}
|
||||
// 最少保留最近 4 条，避免上下文过短
|
||||
if len(kept) < 4 && len(msgs) > 4 {
|
||||
kept = msgs[len(msgs)-4:]
|
||||
}
|
||||
result := append([]llm.Message{msgs[0]}, kept...)
|
||||
return ensureToolMessageIntegrity(result)
|
||||
}
|
||||
|
||||
// ensureToolMessageIntegrity 确保 tool 消息前面有对应 assistant(tool_calls) // Logan.Li
|
||||
func ensureToolMessageIntegrity(msgs []llm.Message) []llm.Message {
|
||||
validIDs := make(map[string]bool)
|
||||
for _, m := range msgs {
|
||||
if m.Role == "assistant" {
|
||||
for _, tc := range m.ToolCalls {
|
||||
validIDs[tc.ID] = true
|
||||
}
|
||||
}
|
||||
}
|
||||
out := make([]llm.Message, 0, len(msgs))
|
||||
for i, m := range msgs {
|
||||
if m.Role == "tool" {
|
||||
if !validIDs[m.ToolCallID] || i == 0 || msgs[i-1].Role != "assistant" {
|
||||
continue
|
||||
}
|
||||
matched := false
|
||||
for _, tc := range msgs[i-1].ToolCalls {
|
||||
if tc.ID == m.ToolCallID {
|
||||
matched = true
|
||||
break
|
||||
}
|
||||
}
|
||||
if !matched {
|
||||
continue
|
||||
}
|
||||
}
|
||||
out = append(out, m)
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
// estimateTokens 粗略 token 估算（中文约 1 字/token，英文约 1/4，折中 rune/3） // Logan.Li
|
||||
func estimateTokens(s string) int {
|
||||
return len([]rune(s))/3 + 1
|
||||
}
|
||||
|
||||
// Close 释放资源 // Logan.Li
|
||||
func (a *Agent) Close() {
|
||||
if a.registry != nil {
|
||||
// Registry 无 Closer 实现（Web 工具无连接池），保留接口位
|
||||
}
|
||||
}
|
||||
Reference in new issue
Block a user